CoChat
Sign Up
Cited by 65Space Satellite Systems and ControlDomain Adaptation and Few-Shot LearningAdvanced Neural Network Applications

Bridging the Domain Gap in Satellite Pose Estimation: A Self-Training Approach Based on Geometrical Constraints

Open access
Wang, Zi, Chen, Minglin, Guo, Yulan, Li, Zhang, Yu, Qifeng·2024·IEEE Transactions on Aerospace and Electronic Systems
DOI 10.1109/taes.2023.3250385
Open this paper in CoChat — free accountOpen PDF

Abstract

Recently, unsupervised domain adaptation in satellite pose estimation has gained increasing attention, aiming at alleviating the annotation cost for training deep models. To this end, we propose a self-training framework based on the domain-agnosticgeometrical constraints. Specifically, we train a neural network to predict the 2D keypoints of a satellite and then use PnP to estimate the pose. The poses of target samples are regarded as latent variables to formulate the task as a minimization problem. Furthermore, we leverage fine-grained segmentation to tackle the information loss issue caused by abstracting the satellite as sparse keypoints. Finally, we iteratively solve the minimization problem in two steps: pseudo-label generation and network training. Experimental results show that our method adapts well to the target domain. Moreover, our method won the 1st place on thesunlamptask of the second international Satellite Pose Estimation Competition.

Verified 2026-10-08 · 1 reference not checked yet

AI snapshot

Written from the paper’s full text.

TL;DR

This paper proposes a self-training unsupervised domain adaptation framework for satellite pose estimation that exploits domain-agnostic geometrical constraints from keypoint projection and mask rendering. It iteratively generates pseudo labels and trains a keypoint-and-mask network, reporting strong SPEED+ results and a 1st-place finish on the sunlamp task of SPEC2021.

Key findings

  • The full framework’s best ablation model \(G_5\) after round 2 achieved lightbox \(E_t = 0.2039\) m, \(E_q = 10.5052^\circ\), \(S = 0.2205\), and sunlamp \(E_t = 0.0963\) m, \(E_q = 4.6960^\circ\), \(S = 0.0988\).
  • Fine-grained segmentation helped: at round 0, model \(G^0_5\) “significantly reduces the estimation error by nearly 30% and 50% in terms of score \(S\) for both domains” compared with the binary-mask variant.
  • In the SPEC2021 comparison, the authors’ entry (lava1302) reported lightbox \(S_t = 0.0464\), \(S_q = 0.1163\), \(S = 0.1627\), and sunlamp \(S_t = 0.0069\), \(S_q = 0.0476\), \(S = 0.0545\); it won 1st place on the sunlamp task and 3rd place on the lightbox task.
  • Source-only validation showed a large domain gap: the authors’ method scored \(S = 0.04\) on synthetic validation but \(S = 2.38\) on lightbox and \(S = 2.87\) on sunlamp.

Methods

  • Experiments use the SPEED+ dataset with the Tango spacecraft: 59,960 labeled synthetic images, 6,740 lightbox images, and 2,791 sunlamp images; the two UDA tasks are synthetic→lightbox and synthetic→sunlamp. Metrics follow SPEC2021, with rotation error \(E_q\), translation error \(E_t\), normalized scores, and thresholds \(\theta_q = 0.169^\circ\), \(\theta_t = 2.173 \times 10^{-3}\).
  • The network predicts 2D keypoint heatmaps and fine-grained masks; poses are estimated via PnP. The backbone is HRFormer-S with 7.8M parameters, with ASPP-based heatmap and mask heads and a discriminator. Training alternates two steps: fix pseudo labels and train the network; fix the network, optimize target poses, and generate pseudo labels, using confidence threshold \(N_{th}=8\).
  • Ablation validation uses manually annotated 100 lightbox and 50 sunlamp images. Training uses 3 rounds: pretraining for 12 epochs at learning rate 0.001 with batch size 5, then 10 epochs at learning rate 0.0005 for later rounds.

Limitations

  • The paper states that only predicted heatmaps are used to generate pseudo poses, while predicted masks are ignored.
  • The paper states that the sparse keypoint representation can make pose estimation less stable when keypoints are invisible due to truncation, low light, or high reflection.
  • The paper states that the projection function \(F\) and rendering function \(R\) are non-differentiable and performed offline; future work is needed for differentiable rendering and online self-training/pose refinement.
  • The text also makes evident that target pose labels are unavailable, so ablation validation relies on manually annotated 100/50 target images, limiting evaluation scale.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/taes.2023.3250385

Cites

67 of 68 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 14 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning

    Models G4 (without geometrical constraints) show degraded performance compared to G5 (with geometrical constraints) in terms of three metrics on both domains.

    “In terms of three metrics on both domains, models G 1 4 and G 2 4 show degraded performance compared to G 1 5 and G 2 5 , respectively.”
    Quote in the paperSubstantial claim
  • Findingown data

    Model G3 benefits from the binary segmentation task.

    “while G 0 3 benefits from the binary segmentation task.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Replacing the transformer-based backbone with HRNet drops pose estimation performance from 0.221/0.099 to 0.253/0.163 on lightbox/sunlamp in terms of score S.

    “the pose estimation performance drops from 0.221/0.099to 0.253/0.163on lightbox/sunlamp in terms of score S.”
    Quote in the paperNot a substantial claim
  • Findingreasoning

    Deep models fully-supervised by synthetic images usually show deteriorated performance when deployed in real scenarios due to the inherent discrepancy between real and synthetic images.

    “Due to the inherent discrepancy between real and synthetic images, the deep models that are fully-supervised by synthetic images usually show deteriorated performance when being deployed in real scenarios.”
    • Also stated in: § C. Source-only Validation · boosted
      “However, all these methods generalize less to the real images.It demonstrates that the domain gap can lead to significant performance degradation.”
    Quote in the paperSubstantial claim
  • Findingown data

    Model G5 shows the highest accuracy by predicting the fine-grained masks.

    “Moreover, G 0 5 shows the highest accuracy by predicting the fine-grained masks.”
    Quote in the paperSubstantial claim
  • May combine two claims. 25
  • Classified as probably not a substantial claim. 10
  • The quote only partly states this claim. 5

  • May combine two claims.

    Findingown data

    Replacing the transformer-based backbone with HRNet reduces running time but increases parameter sizes and memory consumption.

    “Although the running time is shorter at the expense of larger parameter sizes and memory consumption”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Model G2 has larger heatmap losses but achieves better pose estimation accuracy than model G0.

    “Although model G 0 2 has larger heatmap losses, it achieves better pose estimation accuracy than model G 0 0 .”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On the sunlamp domain, our approach outperforms all other methods in terms of translation and rotation scores, taking the 1st place in the challenge.

    “On the sunlamp domain, our approach outperforms all other methods in terms of translation and rotation scores, taking the 1st place in the challenge.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Model G0 overfits more to source samples and generalizes less to target images, because it is only trained using source data.

    “On each domain, model G 0 0 overfits more to source samples and generalizes less to target images, because it is only trained using source data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Our method and HigherHRNet+EPnP obtain the best performance on the synthetic images, illustrating that heatmap-based keypoints regression is capable of fitting the training data.

    “Our method and HigherHRNet [14]+EPnP [14] obtain the best performance on the synthetic images, illustrating that heatmap-based keypoints regression is capable of fitting the training data.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [14] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). May combine two claims.

    Findingnone given

    Using the proposed self-training framework, model G5 achieves the best pose estimation results in all metrics and on all tasks.

    “Moreover, using the proposed self-training framework, model G 2 5 achieves the best pose estimation results in all metrics and on all tasks.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On the lightbox domain, our approach shows competitive performance and has won the 3rd place in the challenge.

    “On the lightbox domain, our approach shows competitive performance and has won the 3rd place in the challenge.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Model G5 performs output-level adaptation by introducing fine-grained segmentation, resulting in the smallest losses and the best pose estimation accuracy on each domain.

    “In contrast, G 0 5 performs output-level adaptation by introducing fine-grained segmentation, resulting in the smallest losses and the best pose estimation accuracy on each domain.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The proposed method won the 1st and the 3rd place on two leaderboards of the second international Satellite Pose Estimation Competition.

    “Moreover, our method won the 1st and the 3rd place on two leaderboards of the second international Satellite Pose Estimation Competition 3 , respectively.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    The proposed method won the 1st place on the sunlamp task of the second international Satellite Pose Estimation Competition.

    “Moreover, our method won the 1st place on the sunlamp task of the second international Satellite Pose Estimation Competition.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingown data

    Model G1 always achieves the largest MSEs on lightbox and sunlamp domains.

    “Note that, model G 0 1 always achieves the largest MSEs on lightbox and sunlamp domains”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingreasoning

    The increase in running time from adding the upsampling layer is only about 0.35%, which is negligible, because the upsampling module and prediction heads are very simple.

    “However, the increase in running time is only about 0.35% (which is negligible) since the upsampling module and the prediction heads are very simple.”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    Adding an upsampling layer to increase feature resolution by a scaling factor 2 significantly improves pose estimation at the expense of increased computation.

    “We observe significant improvements of pose estimation at the expense of increased computation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Fine-grained segmentation can effectively improve domain adaptation performance, and thus prevent the optimization of Eq. (11) converging at local optima.

    “It illustrates that the fine-grained segmentation can effectively improve the domain adaptation performance, and thus prevent the optimization of Eq. ( 11) converging at local optima.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Fine-grained masks provide rich dense descriptions with domain-agnostic structural context.

    “To tackle the above problem, we note that the fine-grained masks in Fig. 1 provide rich dense descriptions with domainagnostic structural context.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    Sparse keypoints stand for only semantic parts of the satellite, and such sparse representation leads to a significant loss of information that hampers knowledge transfer across domains.

    “Such sparse representation leads to a significant loss of information, which hampers knowledge transfer across domains.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    A smaller N_th encourages more estimated poses to be used to generate pseudo labels for model training.

    “A smaller N th encourages more estimated poses to be used to generate pseudo labels for model training.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    Our model can handle intense imaging noise, complicated illuminations, surface reflection, and pose variations.

    “Note that, our model can handle intense imaging noise, complicated illuminations, surface reflection, and pose variations.”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Interpretationreasoning

    With geometrical constraints, models G5 are trained using more clean and more accurate pseudo labels and thus show superior performance.

    “With the geometrical constraints, models G 1 5 and G 2 5 are trained using more clean and more accurate pseudo labels and thus show superior performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The heatmap-based representation of sparse keypoints introduces severe information loss, making the optimization process prone to converging at local optima and leading to deteriorated performance.

    “However, the heatmapbased representation of sparse keypoints introduces severe information loss.Consequently, the optimization process is prone to converging at local optima, leading to deteriorated performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    Since annotations for lightbox and sunlamp domains are not released, there are two UDA tasks with respect to SPEED+: synthetic→lightbox and synthetic→sunlamp.

    “Since the annotations for lightbox and sunlamp domains are not released, there are two UDA tasks with respect to SPEED+, including synthetic→lightbox and synthetic→sunlamp.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Previous satellite pose estimation approaches represent a satellite as a set of 2D keypoints and then estimate the satellite pose using PnP algorithms.

    “Meanwhile, previous satellite pose estimation approaches [12], [15] represent a satellite as a set of 2D keypoints, and then estimate the satellite pose using the perspective-n-point (PnP) algorithms [27].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [12] source

      Not checked in the free run.

    • cites [15] source

      Not checked in the free run.

    • cites [27] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The mask head in model G5 provides more fine-grained predictions and thus effectively enhances structural constraints and contextual information.

    “The key reason is that the mask head in model G 5 provides more fine-grained predictions and thus effectively enhances structural constraints and contextual information.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The SPEED+ dataset comprises images of the Tango spacecraft from the PRISMA mission, consisting of three distinct domains: synthetic, lightbox, and sunlamp.

    “The SPEED+ [22] dataset comprises images of the Tango spacecraft from the PRISMA [49] mission, consisting of three distinct domains, i.e., synthetic, lightbox, and sunlamp.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [22] source

      Not checked in the free run.

    • cites [49] DAmico, 2014, Pose estimation of an uncooperative spacecraft from actual space imagery, Int. J. Space Sci. Eng. source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Leveraging fine-grained segmentation as an auxiliary task has positive impacts on 2D keypoints regression, resulting in better performance of satellite pose estimation across domains.

    “Therefore, leveraging fine-grained segmentation as an auxiliary task has positive impacts on 2D keypoints regression, resulting in better performance of satellite pose estimate across domains.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The fine-grained mask provides dense representation, while the keypoints heatmap is sparse and leads to significant information loss.

    “The fine-grained mask provides dense representation, while the keypoints heatmap is sparse and leads to significant information loss.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The rendering function R provides the same geometrical constraints on source and target samples because the 3D mesh of the satellite and the camera parameters are the same in different domains.

    “Similar to the projection function F, the rendering function R also provides the same geometrical constraints on source and target samples.This is because the 3D mesh of the satellite and the camera parameters are the same in different domains.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Interpretationown data

    The performance degradation of models without geometrical constraints can be ascribed to annotation noise in pseudo labels.

    “The performance degradation can be ascribed to the annotation noise in pseudo labels.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisereasoning

    Due to illumination variations and material discrepancies, the texture of the satellite is not domain-agnostic, while the geometrical model is the same for the source and target samples.

    “However, due to illumination variations and material discrepancies, the texture of satellite is not domain-agnostic, while the geometrical model are the same for the source and target samples.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Only the predicted heatmaps are used to generate pseudo poses, while the predicted masks are ignored.

    “One apparent limitation of our method is that only the predicted heatmaps are used to generate pseudo poses in Sec.III-E, while the predicted masks are ignored.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Interpretationown data

    The proposed method achieves superior performance in satellite pose estimation.

    “The experimental results demonstrate that our method achieves superior performance.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    The satellite is represented sparsely using a set of keypoints.

    “Another limitation would be the sparse representation of the satellite using a set of keypoints.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    The implementations of the projection function F and the rendering function R are non-differentiable and performed offline.

    “The third limitation of this work is that the implementations of the projection function F and the rendering function R are non-differentiable and performed offline.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Pose estimation could be less stable when keypoints are invisible due to truncation, low light, or high reflection.

    “Note that the pose estimation could be less stable when these keypoints are invisible due to truncation, low light, or high reflection.”
    Quote in the paperNot a substantial claim
34 of 34

  • The authors state this as an assumption.

    The poses in the source and target domains are sampled from the same distribution.

    “We assume that the poses in the source and target domains are sampled from the same distribution.”

  • The authors state this as an assumption.

    The texture-less 3D mesh M of the satellite is given.

    “We assume the texture-less 3D mesh M of the satellite is given.”

  • The paper relies on this without support.

    The geometrical constraints used in the self-training framework are domain-agnostic.

    “This paper explores the domain-agnostic geometrical constraints to achieve unsupervised domain adaptation in satellite pose estimation.”

  • The authors state this as an assumption.

    The target poses can be treated as latent variables in the self-training framework.

    “The task is formulated as a minimization problem in a self-training framework by taking the target poses as latent variables.”
4 of 4

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Spacecraft pose estimation networks are exclusively trained on synthetic images because acquiring large sets of real images is difficult.

    “Due to the difficulty of acquiring large sets of real images, spacecraft pose estimation networks are exclusively trained on synthetic ones.”

    From Domain Generalization for In-Orbit 6D Pose Estimation (2025) · cites this paper as [16]

    The Introduction says that collecting and annotating real spacecraft images is difficult and that recent deep models are therefore trained on synthetic images. But “exclusively” overstates this: the Conclusion describes the paper’s own self-training method, which also uses target-domain images during training. The indexed passages do not provide page numbers. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • Synthetic spacecraft images do not capture the illumination conditions encountered in orbit.

    “because those images do not capture the illumination conditions encountered in orbit”

    From Domain Generalization for In-Orbit 6D Pose Estimation (2025) · cites this paper as [16]

    The Introduction (page unavailable) reports a gap between synthetic and real spacecraft images and identifies challenging illumination as a problem. The Dataset and Metrics section (page unavailable) describes lightbox and sunlamp images designed to approximate Earth light and sunlight, but does not establish that synthetic images fail to capture illumination conditions encountered in orbit. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • Both HRNet and Lite-HRNet have been used in spacecraft pose estimation pipelines.

    “Both HRNet [25] and Lite-HRNet [69] have been used in spacecraft pose estimation pipelines [16, 40].”

    From Domain Generalization for In-Orbit 6D Pose Estimation (2025) · cites this paper as [16]

    The Runtime Performance section reports replacing the pose-estimation pipeline’s backbone with HRNet and gives results for HRNet-W32 in Table IV. Our copy does not report using Lite-HRNet, so it supports only the HRNet part of the claim. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

  • Many researchers have adopted approaches that combine keypoints detection with existing Perspective-n-Point (PnP) pose estimation solvers.

    “Consequently, many researchers [8,9,11,14-16] have adopted approaches that combine keypoints detection with existing Perspective-n-Point (PnP) pose estimation solvers.”

    From Pose Estimation for Cross-Domain Non-Cooperative Spacecraft Based on Spatial-Aware Keypoints Regression (2024) · cites this paper as [11]

  • Pose estimation networks trained on synthetic images face a domain gap and do not generalize to real images.

    “Due to these mismatches between the synthetic and real domains, those solutions encounter significant generalization issues.”

    From Domain Generalization for In-Orbit 6D Pose Estimation (2025) · cites this paper as [16]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Table IV labels degree-valued errors as radians

    Minor

    Table IV labels its rotation-error columns in radians, while Table II labels the corresponding errors in degrees. The matching sunlamp values, 4.696 and 4.6960, show that Table IV has not converted those values to radians.

  • Adversarial training does not halve scores in every round

    Minor

    The claimed halving does not hold for both domains in every round. For example, at round 0 the sunlamp score decreases from 0.8651 for G1 to 0.6576 for G2, substantially less than a halving.

  • Geometrical-constraint model does not improve every metric

    Minor

    The text says G1_4 is worse than G1_5 in all three metrics on both domains. In Table II, G1_4 has a slightly lower lightbox translation error at round 1: 0.2342 m versus 0.2346 m.

  • Target joint distribution uses the source-image conditional

    Minor

    Equation (3) factorizes the target-image joint distribution using the conditional probability of a source image. Equation (2) explicitly distinguishes the source- and target-image conditionals, so the stated factorization does not represent the target joint distribution.

4 of 4

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Buslaev, 2020, Albumentations: Fast and flexible image augmentations, 11

Cited by

Sorted by citations

Related

Authors
Wang, Z., Chen, M., Guo, Y., Li, Z., & Yu, Q.
Affiliations
  • National University of Defense Technology (CN)
  • Sun Yat-sen University (CN)
ORCID
Venue
IEEE Transactions on Aerospace and Electronic Systems
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2024-06
Volume
60
Issue
3
Pages
2500-2514
ISSN
  • 0018-9251 (ISSN-L)
  • 1557-9603
  • 2371-9877
DOI
10.1109/taes.2023.3250385
OpenAlex
W4322731048
Semantic Scholar
8bd40fbd7e99cc01e05f44e6d45ee19d75664635
Cited by
64
References
65
Citations by year
2026: 11 · 2025: 28 · 2024: 20 · 2023: 5
FWCI
7.80
Type
journal article
Language
en
Open access
Licence
Locations
  • IEEE Transactions on Aerospace and Electronic Systems (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
National Natural Science Foundation of China (61801491, 61972435, U20A20185)
Topics
  • Physical Sciences › Engineering › Aerospace Engineering › Space Satellite Systems and Control
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
satellite pose estimation, unsupervised domain adaptation, self-training, geometric constraints, 2D keypoint detection, pseudo-labeling, fine-grained segmentation, satellite imagery, annotation cost reduction

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review